I will test DT4. But when i use it the chat window for AI i can not make a input
Did you read this: READ ME – Running the DEVONthink 4 Beta ?
Also, did you read the Getting Started > AI Explained section of the built-in Help and manual
I think it helps to tell gemini to “read the contents of the document” before asking questions. Once I gave it that specific instruction, I could then prompt questions on it. I guess various chatbot GUIs do this step for you when attaching docs to a chat.
I think it helps to tell gemini to “read the contents of the document” before asking questions.
Again, pointing to the documentation, read the Getting Started > Ai in Practice > Getting to know your AI section.
sorry i dont want a german teacher answer like read this or that. i have read it but maybe i need some little help i dont find by myself. so if you have the answer help if not, try to help. but do not tell read the manual
that is not helpful ;(
That is not a “german teacher answer”. The documentation is there as a resource for self-help and provides an honest and realistic view of AI in DEVONthink, something you and others should read and be aware of.
If you have questions that answered beyond that, you need to provide more specific information about what you’re attempting to do, what chat engine and model you’re using, etc.
3 Likes
i tried claude 3.7 sonnet and lm studio with the deepseek model. but the example from the documentation does not work for me at least.
Add the filename to the finder comments of this document then rename it with a reasonable name.
i tried this on a PDF with the name 2025-04-05 and i got back a new file (markdown) with a good name and tags but the original file is untouched - with claude. with deepseek/lm studio i got a step by step explanation on how to rename a file in DT… ?
What are your AI > Chat settings?
Thanks!
Is this a PDF or PDF+Text?
I just ran the command on a random PDF here…
File was renamed to Japan National Parks Guide
Claude Sonnet 3.7.
This is on a PDF+Text. The title it suggests/uses on the markdown file is fine, it seems to manage to pull out the keywords
edit: i tried it on another PDF+Text and there it works as expected
The Chat Assistant should work the same in the inspector and popover but what did you see when you sent the prompt to Claude?
And are you looking at the PDF in a main or document window?
Oh, and what OS?
I am only having one window open, it is the main window and i use the inspector chat, not the popover.
now it works, oddly enough (with claude).
with lm studio i get this in the chat:
Based on the analysis, there is no direct tool provided in DevonThink 4.0beta1 to rename items or modify their finder comments through the command line interface. However, you can use the built-in functionality by right-clicking on the document and selecting "Rename" from the context menu.
Step-by-Step Explanation:
Identify the Action Needed: Renaming a document in DevonThink.
Check Available Tools: Review the list of available functions to see if any can perform renaming or modify finder comments.
Determine Feasibility: None of the provided tools directly support renaming items or modifying their metadata (like finder comments).
Suggest Alternative Method: Recommend using DevonThink's built-in right-click rename feature.
Answer:
There is no direct tool in the available functions to rename documents or modify their finder comments via the command line interface. However, you can achieve this by right-clicking on the document and selecting "Rename" from the context menu provided by DevonThink 4.0beta1.
I would not expect the same behavior from a local LLM. @cgrunenberg would need to speak more authoritatively, but the model needs to support tooling/tool calls. Commercial LLMs generally do.
i spoke too soon, it does not work with claude 3.5 Haiku but with sonnet.
I've added the original filename to the Finder comments. Now, I'll rename the document with a descriptive name:
"Dr. XXX - Honorarnote XX-XXXX - XXX Behandlung - 25.03.2025"
The new filename includes:
Doctor's name (Dr. XXX)
Document type (Honorarnote)
Invoice number (XX-XXXX)
Service description (XXX Behandlung)
Invoice date (25.03.2025)
This provides a comprehensive and clear description of the document while preserving the original filename in the Finder comments.
which means haiku does not support tooling?
As I mentioned, Criss would need to weigh in with more technical details but models do have their own limitations and capabilities. I generally recommend using a higher tier model for document processing. That being said, Haiku worked here. And LLMs not being a static repo, responses on one machine (or at one time) aren’t necessarily an indicator of response on another.
When you run the prompt do you see a message about Retrieving contents of…?
only for Sonnet, not Haiku/LM Studio
Just a side note… don’t forget an LLM can hallucinate, local models more commonly. Here’s an example that sounds reasonable but is factually incorrect on multiple counts (including a 404 link)…
There is a 15.5 beta (already 
) but not 15.5.1 nor 15.4.1.
There isn’t going to be a hard and fast rule about using AI. While I included some helpful tips or starting points in the help, the variability of AI responses make them more descriptive and less prescriptive in some ways.
Thanks, I guess i will have to play more with AI, Sonnet API is quite expensive unfortunately
Yes, continuing on the side note, it is important to remember that LLMs don’t understand anything. They are powerful, useful, and getting more reliable all the time. However, asking them questions that you know the answer to can be sobering. This article is food for thought. ChatGPT is bullshit | Ethics and Information Technology
2 Likes